I have been working in the field of statistical data analysis for the past 20 years, seeking to combine applied mathematics and probability theory with computer science. Nowadays, such a combination is often referred to as "data science", and my primary area of focus is "machine learning", extending also to "artificial intelligence".
Historically, my research interests have mainly centred on the theory and application of probabilistic models for machine learning and statistical pattern recognition, emphasising the use of Bayesian inference. These strands led me to develop the idea of "sparse Bayesian" learning models and to the creation of the relevance vector machine.
I've also spent considerable time working in neural computing, data visualisation, exploratory data analysis, topographic mapping and kernel models. Many of my published papers in these areas are available here on this site.
Application fields I have tackled include bioinformatics, sporting performance analysis, information retrieval, risk assessment, automotive modelling, image processing, handwriting recognition and interactive entertainment.
Read on for an expanded summary of what I've been doing for the last few years (in reverse chronological order), including some relevant further links where useful. I also maintain an entry on LinkedIn.
University of Bath
In April 2016, I joined the Institute for Mathematical Innovation (IMI) at the University of Bath as Professor of Data Science.
FeaturespaceFrom 2013 to 2015, I was Director of Science at Featurespace, a Cambridge (U.K.) based start-up specialising in adaptive behavioural analytics for business. I built a system there for detecting payment card fraud which, at time of writing, appears to be gaining some significant commercial traction.
Vector AnomalyAfter Microsoft, from 2006 to 2012, I was Director of Research at Vector Anomaly, a statistical consultancy which I set up in conjunction with Mark Hatton. We provided advanced statistical data analysis services across a range of business domains, building some interesting software applications for sports performance analysis along the way. Mark continues to keep the business going, developing software for iOS.
From 1998 to 2006, I was a researcher with Microsoft Research (MSR), in Cambridge (UK). I worked primarily on the development and application of probabilistic models in machine learning, as well as many of the other topics mentioned above. In particular, I concentrated on advancing Bayesian methodology for efficient automated prediction and worked on implementations of sparse learning models for regression and classification tasks. This thread of research resulted in the concept of "sparse Bayesian learning" and the popular relevance vector machine.
Latterly at Microsoft, I concentrated on applying machine learning technology to video games, with the aim of both improving existing "artificial intelligence" systems and creating innovative game-play mechanisms which exploit learning techniques. The principal fruit of this research was the Drivatar learning AI system for racing simulation. I developed this (along with Mark Hatton) for the successful Microsoft Xbox title "Forza Motorsport", which shipped in May 2005. It was intriguing to see that this work was resurrected for the recent "Forza Motorsport 5" Xbox One launch title — it must have been ahead of its time :-)
NCRG, Aston University
From 1992 to 1998, I was a member of the Neural Computing Research Group at Aston University. From that time, the following links may still be of some interest:
Netlab "neural network software". Netlab (to which I am a very minor contributor) runs under Matlab® and in addition to neural nets, also contains many other useful modelling tools. For what it's worth, I highly recommend both the free software and its accompanying text book.
- Probabilistic PCA, and mixtures thereof. I have received many requests over the last few years concerning the availability of code to implement mixtures of probabilistic principal component analysers [Paper abstract and download]. I've never been able to find time to produce a version of my own, but I did assist with the implementation in Matlab® that is available in the aforementioned Netlab toolbox.
PhiVis is a Matlab® toolbox for visualisation of multivariate continuous data sets, exploiting hierarchical mixtures of probabilistic principal component projections. The software appears to no longer be available from Aston, but see the related journal article for further details.
Before Time Began
I moved to the NCRG subsequent to completing a one-year M.Sc. course in "Knowledge-Based Systems" at Edinburgh University (1991-2). My masters thesis was on "Digital Filtering with Recurrent Neural Networks".
Prior to that, I obtained a B.Eng (honours) degree in Electrical and Electronic Engineering at Bristol University (1987-90), and spent a lot of time playing football.
In 1986-87 I worked at G.E.C. Industrial Controls (Kidsgrove, Stoke-on-Trent) for a year as part of a "sandwich" student placement.
I attended Newcastle High School (later Newcastle-under-Lyme School), Staffordshire, from 1979-86.
Surely, by now, no-one can be reading this any more ...